portfolio/labs/linux/setup/docs/nvidia.md

NVIDIA Tooling

Diagnostic-only default

The normal NVIDIA profile installs nvtop, clinfo, and PCI utilities. It does not install or select a driver:

sudo ./install.sh --nvidia-tools

Review hardware and current module state:

lspci -nn | grep -i nvidia
nvidia-smi
dkms status
mokutil --sb-state

Explicit driver installation

Install only a reviewed Ubuntu driver package version:

sudo ./install.sh --install-nvidia-driver 550

The numeric value maps directly to nvidia-driver-VERSION. The profile refuses an unavailable package. Reboot after installation, then validate nvidia-smi, kernel logs, DKMS state, and application behavior.

Selection considerations

• GPU generation and supported driver branch.
• Ubuntu release, kernel, and HWE stack.
• Secure Boot module enrollment.
• CUDA or application compatibility.
• Docker NVIDIA Container Toolkit requirements.
• Whether the device will be bound to VFIO instead of the host driver.

Troubleshooting

journalctl -k | grep -Ei 'nvidia|nouveau|NVRM'
lsmod | grep -E 'nvidia|nouveau'
dkms status
apt-cache policy 'nvidia-driver-*'

Driver rollback is environment-specific and is not automated. Preserve console access and a known-good kernel before changing GPU or Secure Boot configuration.